FAR-GAN: Color-controlled Fashion Apparel Regeneration

Abstract

Automatic fashion apparel regeneration is an important aspect for the e-commerce retailers to provide an opportunity to preview the selected dress in the desired color. This helps in improving customer satisfaction and sales. In this work, we propose FAR-GAN, a fashion apparel synthesis tool with explicit control on color. The proposed approach augments the features from the fashion apparel and its edge-map in a two-step encoding process to extract the style information. This information is controlled with the target color embedding information in the decoder. To control the color of the synthesized apparel image, we have proposed the color consistency loss. Overall, the network can be trained end-to-end without incorporating any complex sub-units and controlling the color of the choice for the synthesized product image. We have conducted extensive experiments and ablation study to showcase the performance of our model compared to several state-of-the-art methodologies. The results reflect improvement in performance and justification of our design choices.